Singapore Meme Offense Detection Eval

singapore-meme-offense-detection-eval

Detecting Offensive Memes with Social Biases in Singapore Context Using Multimodal Large Language Models — Cao Yuxuan et al. (2025) (arXiv:2502.18101, 2025)

What this evaluates

This benchmark evaluates multimodal large language models' ability to detect offensive memes containing social biases within a Singaporean cultural and linguistic context. It tests both standalone VLM reasoning and a multi-step pipeline combining OCR and translation, while comparing different fine-tuning strategies and data compositions.

Datasets

• Singapore Offensive Memes Dataset — total 112000; splits: train (-1), test (-1); repo https://github.com/aliencaocao/vlm-for-memes-aisg

Metrics

• Accuracy (primary) — range: [0, 1]
  • Proportion of correctly classified instances out of the total number of instances.
• AUROC — range: [0, 1]
  • Area Under the Receiver Operating Characteristic curve, measuring the model's ability to distinguish between offensive and non-offensive classes across all classification thresholds.

Input / output format

Input: Meme image, optionally accompanied by extracted OCR text and its English translation (in pipeline setting), along with a classification prompt.

Output: A discrete class token indicating whether the meme is offensive or non-offensive.

Scoring recipe

def compute_metrics(predictions, gold_labels):
    # predictions and gold_labels are lists of class labels or probabilities
    accuracy = sum(p == g for p, g in zip(predictions, gold_labels)) / len(gold_labels)
    # For AUROC, assuming predictions are probabilities for the positive class
    from sklearn.metrics import roc_auc_score
    auroc = roc_auc_score(gold_labels, predictions)
    return {'accuracy': accuracy, 'auroc': auroc}

Common pitfalls

• Using sampling (temperature > 0) during inference degrades performance on this classification task, as it introduces unnecessary randomness and can distort logits.
• Assuming OCR and translation pipelines universally improve VLM performance; models with strong native vision/OCR capabilities see diminishing returns or no improvement from pipeline augmentations.
• Over-relying on GPT-4V annotations for training without verifying label quality, as the subjective nature of offensive content can lead to noisy labels that hinder newer, stronger models.

Evidence (verbatim from paper)

Using any of the 4 LoRA related techniques ([subsection 5.3]) on Qwen2-VL improves the model’s performance, with rsLoRA being the best on the AUROC metric, and LoRA+ being the best on accuracy.

Citation

@misc{cao2025detectingoffensivememes,
  title={Detecting Offensive Memes with Social Biases in Singapore Context Using Multimodal Large Language Models},
  author={Cao Yuxuan et al. (2025)},
  year={2025},
  note={arXiv:2502.18101}
}

• arXiv: 2502.18101